Related papers: Scheming in the wild: detecting real-world AI sche…
Online Social Networks (OSNs), such as Facebook, provide users with tools to share information along with a set of privacy controls preferences to regulate the spread of information. Current privacy controls are efficient to protect content…
Recent work has demonstrated the plausibility of frontier AI models scheming -- knowingly and covertly pursuing an objective misaligned with its developer's intentions. Such behavior could be very hard to detect, and if present in future…
We sketch how developers of frontier AI systems could construct a structured rationale -- a 'safety case' -- that an AI system is unlikely to cause catastrophic outcomes through scheming. Scheming is a potential threat model where AI…
Highly capable AI systems could secretly pursue misaligned goals -- what we call "scheming". Because a scheming AI would deliberately try to hide its misaligned goals and actions, measuring and mitigating scheming requires different…
Frontier models are increasingly trained and deployed as autonomous agent. One safety concern is that AI agents might covertly pursue misaligned goals, hiding their true capabilities and objectives - also known as scheming. We study whether…
Small businesses need vulnerability assessments to identify and mitigate cyber risks. Cybersecurity clinics provide a solution by offering students hands-on experience while delivering free vulnerability assessments to local organizations.…
Open Source Intelligence (OSINT) refers to intelligence efforts based on freely available data. It has become a frequent topic of conversation on social media, where private users or networks can share their findings. Such data is highly…
Facebook represents the current de-facto choice for social media, changing the nature of social relationships. The increasing amount of personal information that runs through this platform publicly exposes user behaviour and social trends,…
Open Source Intelligence (OSINT) investigations, which rely entirely on publicly available data such as social media, play an increasingly important role in solving crimes and holding governments accountable. The growing volume of data and…
Intent Detection systems in the real world are exposed to complexities of imbalanced datasets containing varying perception of intent, unintended correlations and domain-specific aberrations. To facilitate benchmarking which can reflect…
This paper examines the role of Open Source Intelligence (OSINT) on Twitter regarding the Russo-Ukrainian war, distinguishing between genuine OSINT and deceptive misinformation efforts, termed "BULLSHINT." Utilizing a dataset spanning from…
As large language models (LLMs) are increasingly deployed as interactive agents, open-ended human-AI interactions can involve deceptive behaviors with serious real-world consequences, yet existing evaluations remain largely…
As frontier language models are increasingly deployed as autonomous agents pursuing complex, long-term objectives, there is increased risk of scheming: agents covertly pursuing misaligned goals. Prior work has focused on showing agents are…
Artificial intelligence systems are now deployed at scale across sectors, accompanied by a growing number of real-world incidents ranging from misinformation and cybercrime to autonomous-system failures. Databases of AI incidents index…
High-risk industries like nuclear and aviation use real-time monitoring to detect dangerous system conditions. Similarly, Large Language Models (LLMs) need monitoring safeguards. We propose a real-time framework to predict harmful AI…
As artificial intelligence (AI) systems become increasingly deployed across the world, they are also increasingly implicated in AI incidents - harm events to individuals and society. As a result, industry, civil society, and governments…
This report examines whether advanced AIs that perform well in training will be doing so in order to gain power later -- a behavior I call "scheming" (also sometimes called "deceptive alignment"). I conclude that scheming is a disturbingly…
We introduce scheming honeypot evaluations, a framework for testing whether models will pursue instrumental goals if given the opportunity. Our scheming honeypot evaluations take the form of coding tasks in Google's alignment research…
Two years after publicly launching the AI Incident Database (AIID) as a collection of harms or near harms produced by AI in the world, a backlog of "issues" that do not meet its incident ingestion criteria have accumulated in its review…
We systematically evaluate the quality of widely used adversarial safety datasets from two perspectives: in isolation and in practice. In isolation, we examine how well these datasets reflect real-world adversarial attacks based on three…